fix(dialect): make extend_sqlglot idempotent

[Sanjays2402]

Closes #5908

Description

_override saved whatever was currently installed under _<name>, so a second extend_sqlglot() call saved the override itself as the "original". The wrapper then called itself and every subsequent type/CAST parse raised RecursionError, with sqlglot's original method lost. This is reachable without calling it twice on purpose: extend_sqlglot() is the first statement of sqlmesh/__init__.py, so if anything later in that import cascade raises, the package is evicted from sys.modules while the patched globals survive, and the retry re-applies the patch.

_override now returns early when the override is already installed.

Test Plan

Added tests/core/test_dialect.py::test_extend_sqlglot_is_idempotent, which fails with RecursionError without the fix and passes with it. tests/core/test_dialect.py is green (160 passed).

Checklist

• I have run make style and fixed any issues
• I have added tests for my changes (if applicable)
• All existing tests pass (make fast-test)
• My commits are signed off (git commit -s) per the DCO

[StuffbyYuki]

@Sanjays2402 Thanks for the PR! Here are a few things (not blockers):

• Also guard generator.UNWRAPPED_INTERVAL_VALUES (still appends every call; TRANSFORMS / WITH_SEPARATED_COMMENTS already have membership guards — called out in #5908).
• Optional assert that a second call doesn’t grow UNWRAPPED_INTERVAL_VALUES (or that __parse_types still points at sqlglot’s original).

Merge ready once ci tests pass

[Sanjays2402]

Thanks @StuffbyYuki — pushed 5f2c7b9.

Added a MacroVar not in generator.UNWRAPPED_INTERVAL_VALUES guard so it matches the TRANSFORMS / WITH_SEPARATED_COMMENTS pattern, and the idempotence test now snapshots Generator.UNWRAPPED_INTERVAL_VALUES before the second extend_sqlglot() call and asserts it's unchanged.

On the 3.9 style job: the 12 failures are all TypeError: unsupported operand type(s) for |: 'type' and 'type' in tests/dbt (bigquery relation/column classes), which is PEP 604 syntax at runtime on 3.9 — unrelated to this diff. tests/core/test_dialect.py is green locally (160 passed).

[StuffbyYuki]

@Sanjays2402 Can you add DCO checks?

[Sanjays2402]

Done — signed off the second commit and force-pushed, head is now fb992b7. 10c1b46 already had the trailer; 132f16b (the UNWRAPPED_INTERVAL_VALUES guard) was missing it, which is what the Commits Check was failing on.

Amended with --trailer only, so the tree is byte-identical to 132f16b (git diff 132f16b HEAD --stat is empty). CI is re-queuing now.